rishikksh20/AdaSpeech

Built on FastSpeech 2, this PyTorch implementation applies utterance-level and phoneme-level acoustic embeddings to improve generalization across speaking styles without requiring condition layer normalization. It preprocesses audio using Montreal Forced Aligner for duration extraction and normalizes prosodic features (F0, energy) per dataset, then trains an encoder-decoder architecture that conditions acoustic predictions on speaker characteristics. Targets single-speaker TTS scenarios where adaptive acoustic modeling enhances synthesis quality beyond base FastSpeech 2.

162 stars. No commits in the last 6 months.